JPS6332600A - Voice detector - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a voice detection device suitable for use in digital communications, voice storage systems, or other voice signal processing.

(Prior Art) In a conventional communication system, two people have a conversation via a two-way telephone line, so that the person on the other end of the conversation listens. As a result, voice signals often flow only in one direction, and the lines in the opposite direction are empty. Furthermore, even in the middle of a speech, there are moments when the audio signal fails at syllable breaks or other places.

In this way, when one line is assigned to each caller,
The probability that a voice signal is flowing through each line is usually less than 1/2. Therefore, by allocating a line to one caller only when voice is present, it becomes possible to handle more calls than the number of lines.

By the way, in order to realize such a multi-line call, it is necessary to detect the section of the voice signal flowing through the line and allocate calls. As a detection method for this audio signal, TASI (Time Assignment Intercollation), which targets analog audio signals, has conventionally been used.
ime ASsignsent 5peech In
terpolation (abbreviation)) and DSI (digital speech
Interturbo 1/Ission ([11g1talSpeec
h Interpolation) law (for example, the document “Transactions of the Institute of Electronics and Communication Engineers 80/7.J63-A (7
) p413-420) are known.

This TASI method is most commonly used in long-distance submarine coaxial cable systems. On the other hand, the DSI method is mainly used in digital satellite communication lines because voice detection and channel allocation control operations are relatively easy.

According to this DSI method, as described above, the DSI method is roughly divided into a voice detection section and a channel allocation control section. In particular, DSI
The specifications of this voice detection section required by the law include short voice detection time, few malfunctions caused by sounds, and correct voice detection.

Conventionally, the voice detection device used in this DSI method,
The presence or absence of speech is determined by detecting the speech waveform, that is, the speech section, using only the electric power ht.

Therefore, it is necessary to add a long hangover time, for example 150 to 250 m5, in order to avoid or eliminate truncations at the beginning of words, mid-speech and endings, etc.6
However, this poses a problem in that a long hangover time is added due to malfunctions caused by noise, and as a result, the gain of the voice detection device is reduced.

Therefore, it is known that thermal noise other than voice usually has a uniform frequency distribution and a Gaussian power distribution, whereas the power distribution of voice is a Laplace distribution and a random frequency distribution. In order to improve the detection ability of voice detection, it is also possible to perform detection using not only the amount of power but also the qualitative difference (energy distribution, frequency distribution, or other suitable amount) between noise and voice. Conceivable.

As a conventionally known method for identifying qualitative differences between speech and noise, a method has been proposed that uses the number of times a speech waveform crosses a zero point within a certain period of time (zero-crossing method). However, since this zero-crossing method identifies whether a certain range of frequencies exists in the audio waveform, it cannot identify the magnitude of the power amount of a certain range of frequencies in the audio waveform. do not have.

In addition, the voice detection device used in voice recognition devices is
Compared to the voice detection device used in the DSI method, the presence or absence of voice is identified by digital signal processing of the voice waveform over a long period of time, so the accuracy is high, but the long detection time makes the DSI method Not applicable.

As described above, since there is no conventional method for identifying the frequency distribution that takes into account the power amount of a voice waveform, conventional voice detection devices have ended up using a voice detection method based only on the power amount.

FIG. 5 is a block diagram illustrating an example of the configuration of a conventional voice detection device used in the DSI method and which performs voice detection based only on electric power.

In FIG. 5, Dl is the input PCM signal, typically sampled at 8 kHz. 50 is an absolute value calculation circuit which removes the positive and negative signs of the input PCM signal D1 and outputs an output D2 representing the absolute value to the first comparator 52 at the next stage. In this first comparator 52, the rice planting T1 having a size set empirically in advance is compared with this output D2, and the absolute value D2 of the input signal DI is larger than twice the threshold value T1. In this case, the output signal D3 is set to +2 and the threshold T1
If the quantity is larger but double, the output signal D3 is set to +1, and if it is less than or equal to the threshold T1, the output D3 is set to -1, and the output D3 to which each value is assigned is set. is supplied to an accumulator 54. In this accumulator 54, the output D that is sent during a certain period of time, for example, 4 ms.
Add the value of 3. The added value D4 from the accumulator 54 is sent to a second comparator 56, where it is compared with a threshold value T2 previously set based on experience. Added value D4 is threshold T2
In this case, the output D5 of the second comparator 56 is set to binary "1" and it is determined that there is audio. On the other hand, if the added value D4 of the accumulator 54 is less than or equal to the threshold T2,
The output p5 of the second comparator 56 is set to "0" and it is determined that the voice has disappeared.

This output D5 is supplied to the Hang-O-8 hour adding circuit 58 at the next stage. In this case, in order to prevent voice endings being cut off, disconnections during calls, etc., the output D5 changes from "1" to "0".
, a hangover time is added at the end of the rlJ state of this output to keep it in the "l" state for a certain period of time, thereby outputting an output D6 with an extended duration. It is configured.

(Problems to be Solved by the Invention) However, since the voice detection device with this conventional configuration uses a method of detecting the voice waveform based on the amount of electric power, it still suffers from the problem described above, that is, due to malfunction due to noise. 8 hours are added to the long hang-o, and therefore the detection time becomes longer than the last voice section, which causes a problem in that the gain of the voice detection device decreases.

In view of the problems of the conventional speech detection device mentioned above, it is an object of the present invention to prevent malfunctions caused by noise by detecting speech sections with high precision, Nakasujdan, ending! It is an object of the present invention to provide a voice detection device which eliminates TJ disconnection, does not cause a decrease in gain, and has excellent voice detection ability.

(Means for Solving the Problems) In order to achieve this objective, the voice detection device of the present invention takes the following measures.

First, this voice detection device includes a first voice detection section that outputs a first voice detection signal that determines a voice waveform, that is, a voice section, based on the amount of power during a certain period of a digital input signal.

Furthermore, a second voice detection section outputs a second voice detection signal that determines the voice waveform, that is, the voice section, based on the power ratio between components of mutually different frequency bands in the same period as before the digital input signal. Equipped with.

Furthermore, a determination unit is provided which determines a voice section according to the urgency of the voice of the digital input signal by logically calculating the first and second voice detection signals.

In this case, it is preferable that the first audio detection unit includes one unit for calculating the power during a certain period of the digital input signal.
It is preferable to provide a power calculation section and a comparator that compares the obtained power with a predetermined threshold value and outputs the comparison result as, for example, a binary voice detection signal.

Preferably, the second audio detection unit includes a plurality of filters each capable of outputting components of different frequency bands of the digital input signal, and calculates the power amount of each of the outputs of these filters in a certain period as described above. Calculate the ratio between the multiple power calculation units and the appropriate power amounts obtained from these main power calculation units, compare these power ratio values with the Lll value, and compare the comparison results with, for example, two power calculation units. It is preferable to provide a comparator that outputs the first voice detection signal as the first voice detection signal.

These filters may be a combination of multiple filters with different appropriate pass frequency bands, such as a low-pass filter, a high-pass filter, a fir-pass filter with an appropriate cutoff frequency, or a band-elimination filter. It can be used.

(Function) As described above, according to the voice detection device of the present invention, the one-sound section detection signal detected based only on the power amount of the input signal and the input signal passed through two or more digital e-wave devices Thereafter, the presence or absence of speech in the input signal is determined based on the speech section detection signal detected based on the power ratio between components of different frequency bands.

As already explained, since speech and 31 sounds are usually qualitatively different, the power ratio between different frequency components of the input signal is a measurement quantity that reflects this qualitative difference between speech and noise. Therefore, detection of speech intervals based on this,
By combining the voice section detection based only on the amount of power, it is possible to eliminate erroneous voice section detection due to noise and to perform voice section detection with high sensitivity.

(Embodiments) Hereinafter, embodiments of the voice detection device of the present invention will be described with reference to the drawings.

It should be noted that the embodiments described below are merely preferred examples, and therefore, the present invention does not apply to the numerical examples illustrated in these embodiments,
It is clear that the present invention is not limited only to the configuration of the processing means, processing procedure, etc., and that various changes and modifications can be made within the scope of the present invention.

FIG. 1 is a block diagram showing the configuration of an embodiment of the voice detection device of the present invention. In the figure, 10 is a first voice detection section, 12 is a second voice detection section, 14 is a determination section, and 16 is a second voice detection section.
Although not necessarily included in this device, is a hangover time adding circuit.

The angle detectors IO and 12 are supplied with a digital input signal DIO, for example, a 14-bit signal obtained by linearly converting a PCM signal sampled at 8 kHz.

Description of First Audio Detection Unit As already explained, the first audio detection unit 10 uses a first audio detection signal that determines the presence or absence of an audio waveform, that is, audio section, based on the amount of power during a certain period of the input signal DIO. D
12.

Therefore, in this embodiment, the first audio detection unit IO preferably includes a power calculation unit 20 that calculates the power during a certain period of the digital input signal DIO, and a power calculation unit 20 that calculates the power in a certain period of the digital input signal DIO, and a power calculation unit 20 that calculates the power in a certain period of the digital input signal DIO, and a A comparator 21 is provided for comparing and outputting the comparison result as, for example, a binary first voice detection signal.

Next, an example of the operation of the first voice detection section 10 will be described with reference to the flowchart of the operation shown in FIG.

When the input signal DIO is input to the power calculation unit 20 (step 1. Hereinafter, the step is indicated by S), calculation processing is performed to calculate the power for each sample (S2), and a certain amount of the calculated power is Regarding the number of samples (usually 32 to 8
Approximately 0 samples, time is approximately 4ms ec ~ 10m
It corresponds to about sec. ) A power sum calculation process is performed to obtain the total sum (S3), and the total power D16 is output to the comparator 21 at the next stage. If the power summation of a certain number of samples has not been completed, the processes of steps S1 to S3 are repeated, and if this process has been completed, the process proceeds to the next process (S4).

Next, this comparator 21 performs a comparison process between f-1(ffiT+o for voice section determination and this total power Dl13), and if the total power DIO is north of the threshold TIO, there is voice. The determination information of binary "1" representing the sound and the determination information of binary "0" representing the absence of sound when it is smaller than the threshold D10 are respectively output as the first sound detection signal (S5). Determine whether the signal D10 has ended, and if it has not ended, step 3
Steps 1 to S5 are repeated until the signal ends (S6).

Note that this threshold value TIO can be set in advance based on experience. Further, the power calculation unit 20 and the comparator 21 are provided with means for performing the corresponding processes described above. Also, in particular, the power calculation unit 20 can preferably be constructed from a signal processor operated by Microsoft.

Description of Second Audio Detection Unit The second audio detection unit 12 also includes a second audio detection unit 12 that determines the audio waveform, that is, audio section, based on the power ratio between components of different frequency bands in the same period as before the digital input signal DIO. It is configured to output a two-voice detection signal D14.

In this embodiment, the second audio detection unit 12 preferably includes a plurality of low-pass filters, such as a first filter 22 and a second filter 23, each of which can output components of different frequency bands of the digital input signal. A high-pass filter is provided. These filters 22 and 23 are second-order digital filters, and are a low-pass filter and a high-pass filter, respectively, having cutoff frequencies at the center of the frequency range used for the PCM signal. Output DI8 of components in different frequency ranges obtained from these filters 22 and 23
．． D20, the amount of electric power D22. D
A plurality of first and second power calculation units 24 and 25 are provided, each of which calculates and outputs 24. Further power gained%
Calculate the ratio value between D22 and D024, compare these power ratio values with a threshold value, and use the comparison result as, for example, a binary second voice detection signal D14 for determining the voice waveform, that is, the voice section. A comparator 2B for output is provided.

First and second filters 22 and 2 used in this example
3 can be a digital filter having a configuration as shown in FIG. 2, for example. In Figure 2, 30 and 31
are the input terminal and the output terminal f. 33 to 36 are adders;
37 to 41 are multipliers, and 42 and 43 are delay elements. Then, from the input terminal 7-30, an adder 33 and a multiplier 3
9, connect to adder 35 and output terminal T-31. Further, the adder 33 and the multiplier 38 are connected to the other input terminal of the adder 33 via a delay element 40, a multiplier 37, and an adder 34. Also, between the delay element 42 and the multiplier 37, on one side, the delay element 43 and the multiplier 38 are connected, and then the adder 34 is connected.
It is connected to the other input terminal r of the adder 35 at the other end via the multiplier 43 and the adder 36. Further, the delay element and the multiplier 38 are connected to the other input terminal of the adder 36 via the multiplier 41.

In this embodiment, the digital filter exhibits a second-order transfer function. This filter takes the input as Vin and the output as Vou, and sets the multiplication coefficients of the multipliers 37 to 41 as bl, b2 . These are appropriately set as ao, al, and a2 using a conventionally known method, and -
If the delayed arrival of delay elements - is denoted as z-l, then the input V
i□ and output■. It can be configured such that the relational expression (1) holds between I and I.

Next, the operation of the second voice detection section will be explained with reference to the flowchart of the operation shown in FIG.

When the input signal D10 is input to the first and second filters 22 and 23 (SIO), components in the low and high frequency ranges are outputted and sent to the power calculation units 24 and 25 in the next stage.
(S11), these power calculation units 2
4 and 25, arithmetic processing is performed to calculate the power for each sample (S12), and among the determined powers, the same constant number of samples as in the case of the first audio detection unit 10 (usually 32 to 80 It takes about 4m5 for a sample.
It corresponds to about ec~10m5ec. ) A power sum calculation process is performed to obtain the total sum, and the total power D22 and D24 are respectively output to the next stage comparator 21 (S13).

If the power summation for a certain number of samples is not completed,
The process of steps S10 to S13 is repeated, and if this process is completed, proceed to the next process (514)
.

Next, in this comparator 21, first, two filters 22 and 23 and their respective power calculation units 24 and 25 are used.
An arithmetic process is performed to calculate the ratio of the total powers D22 and D24 obtained through (S15). If the input signal Dlo is noise, the power ratio will be approximately 1 because it is uniformly distributed; on the other hand, When the input signal DIO is audio, the power ratio takes a value other than 1 because the frequency distribution differs depending on the type of audio.

Next, a comparison process is performed between this power ratio and a threshold value TI2 previously set based on experience for voice section determination (SI
6). For example, if the internal detective TI2 is 1.1 and 0.8, when the power ratio ≧1.1 or ≦0.9, the judgment information of binary “1” representing voice I1 and 1.1≧'+h Force ratio≧
In the case of 0.9, determination information of a z value of "0" representing noise and no voice is output as the first voice detection signal D14. Then, it is determined whether the input signal DIO has dropped, and if it has not finished, steps 810 to S
Step 18 is repeated until the signal ends (S17).

Note that the first and second filters 22 and 23, the power calculation section 2
4 and 25, and the comparator 26 are provided with means for performing the corresponding processing.

Further, when such a low-pass filter and a high-pass filter are used, it is particularly suitable for detecting consonants having large energy in a frequency range of 3 kHz or more and vowels having large energy in a frequency range of 1 kHz or less.

Description of Determination Section and Hangover Time Adding Circuit Since these first and second voice detection signals D12 and D14 are binary signals, the voice section of these digital input signals is determined by logical operations. In this embodiment, the 'legged part 14
is constructed using an ordinary logical sum (OR) circuit. Therefore, when either or both of the first and second voice detection units IO and 12 are signals DI2 and D14 of “1” indicating voice detection, the determination unit 14 detects a voice of “1”. Outputs the signal D26 and outputs this "1" signal! Indicates that the continuous period is a voice section. In this case,
When the sound level is dog, the first sound detection unit 10 is mainly heard, and when the sound level is close to the noise level, the second sound detection unit 12 is mainly heard.

Then, when this voice detection signal D2B changes from rlJ to rQJ, the hangover time adding circuit 1
8, detect this change and keep it in the “1” state for an appropriate amount of time! Continuing one line to prevent endings, interruptions, etc.

The configuration of the voice detection device of the present invention described above can be easily formed using software technology or a combination of software technology and hardware technology.

Further, since the power calculation section described above can be configured by a communication signal processor operated by Microsoft, the present invention is suitable for application to a communication device such as a DSI or a voice detection device of a voice storage system.

(Effects of the Invention) As is clear from the above description, according to the voice detection device of the present invention, in voice detection, the electric power of the input signal is used in conjunction with the qualitative difference between voice and noise to detect the voice. Since the presence or absence is detected and the detection method is a detection method that takes both frequency and power into consideration, erroneous voice detection due to noise can be eliminated and voice detection can be performed with high sensitivity.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the voice detection device of the present invention, FIG. 2 is a block diagram showing an example configuration of a digital filter used in the voice detection device of FIG. 1, and FIG.・A flowchart of the operation to explain the voice detection section. FIG. 4 is a flowchart of the operation to explain the first and second voice detection sections. FIG. 5 is a block diagram showing an example of the configuration of a conventional voice detection device. lO...First audio detection unit, 12...Second audio detection unit 14...Judgment unit 16...Hangover time addition circuit 20.24.25...Power calculation unit, 21.28...・
Comparator 22...first filter (first filter) 23...
・Second filter (second filter) 30...input end f,
31... Output terminals 33-3B... Adder,
37-41... Multipliers 42, 43... Delay elements. Patent Applicant: Oki Electric Industry Co., Ltd. No.-@P Kisokuyo-9Q

Claims (3)

[Claims] (1) In a voice detection device that detects a voice section from a digital input signal, a first voice detection section that outputs a first voice detection signal that determines a voice section based on the amount of power during a certain period of the digital input signal; a second voice detection unit that outputs a second voice detection signal that determines a voice section based on a power ratio between components of different frequency bands in the certain period of the digital input signal; and the first and second voice detection. A voice detection device comprising: a determination unit that determines a voice section of the digital input signal by performing a logical operation on the signal. (2) The first audio detection unit includes a power calculation unit for calculating the power in a certain period of the digital input signal, and a first audio detection unit that compares the obtained power with a predetermined threshold and uses the comparison result as the first audio detection unit. The voice detection device according to claim 1, further comprising a comparator that outputs the signal as a signal. (3) The second audio detection unit includes a plurality of filters each capable of outputting components of different frequency bands of the digital input signal, and a plurality of power filters each capable of calculating the amount of power of the output of these filters in the certain period. A patent characterized in that it comprises a calculation unit and a comparator that calculates a power ratio between components of different frequency bands, compares the power ratio with a threshold value, and outputs the comparison result as a second audio detection signal. A voice detection device according to claim 1.